Three dimension(3D) reconstruction is an important research field in computer vision, which is widely used in virtual reality, digital entertainment, 3D electronic map, medical imaging, etc. Generally,the process of 3D reconstruction includes three steps: image matching, camera calibration, and recovery of 3D surface structure. Among these steps, image matching is the most important and critical problem. For outdoor building scenes, the textures are generally repetitive and locally insufficient, this brings difficulties to image matching and quasi-dense matching. This thesis focuses on how to resolve these difficulties and the main contributions include: 1.At least quasi-dense matching is required for 3D reconstruction in order to have better 3D visualization. We discuss the main frames of quasi-dense matching algorithms. In particular, we analyze a previous quasi-dense narrow baseline matching algorithm and a previous quasi-dense wide baseline matching algorithm, then we give their performance differences including different feature detectors, different propagation strategies such as corresponding neighborhood window selection. 2.A novel quasi-dense matching algorithm for building scenes with repetitive and locally insufficient textures is proposed. A new region detector MSERDoG(maximally stable extremal regions on difference of gaussian space) is used to obtain more initial seed matches, then affine propagation with ASW(adaptive support-weight) score is used to eliminate match ambiguity. In addition, matches on borders become more precise. 3.We discuss some classical problems in existing quasi-dense matching algorithms, like edge matching, adjustment of propagated matches, controllability of the propagation process, propagations from two above images. Base on building scenes, we provide some thoughts and possible solutions for these problems.
修改评论